Atlas

Models

← All models

Nemotron 3 Ultra 550B A55B

NVIDIA · 2026-06-04 · 560.5B parameters

Nemotron 3 Ultra 550B A55B is NVIDIA's open-weight, post-trained frontier reasoning model in the 550-billion-parameter class, with a publicly reported 55 billion active parameters per token, released on June 4, 2026. It uses a hybrid LatentMoE architecture combining Mamba-2, MoE, and attention layers with MTP, supports configurable reasoning, and has a documented maximum context of 1,048,576 tokens even though the checkpoint configuration defaults to 262,144. NVIDIA pre-trained it with an NVFP4 recipe and post-trained it with SFT, reinforcement learning, and MOPD for coding, math, instruction following, tool use, and long-context agentic workflows; the BF16 checkpoint is released under OpenMDW-1.1.

Benchmark scores

BenchmarkScore
A-Fantasia73.7
A-Fantasia Backwards Spelling89.0
A-Fantasia Chess59.0
A-Fantasia Cube Rotation73.0
AA-Briefcase Elo874
AA-LCR67.0
AA-Omniscience Index-0.8
AIME 202694.2
Artificial Analysis Agentic Index27.4
Artificial Analysis Coding Index49.3
Artificial Analysis Intelligence Index37.8
Artificial Analysis Omniscience Accuracy21.6
Artificial Analysis Omniscience Hallucination Rate28.5
Artificial Analysis Openness Index83.3
AutomationBench-AA5.7
BrowseComp63.0
BullshitBench v2 Leaderboard49.0
Capability148.5
ComplexConstraints17.8
CorpFin v265.5
CorpFin v2 - Exact Pages64.3
CorpFin v2 - Max Fitting Context67.1
CorpFin v2 - Shared Max Context64.9
CRED v0.9271.5
CRED v0.92 - Data integrity85.0
CRED v0.92 - Execution100.0
CRED v0.92 - Paper–code56.7
CRED v0.92 - Prose–table76.0
CritPt3.1
Design Arena Overall Code Categories1156
Loading Atlas data…